Also in This Week's Premium Deep Dives

→ GPT-5.4 Native Computer Use: 5 Automation Workflows

→ Google Gemini Life Hub: Setup & Privacy Guide

→ Pro Tip: Automate Browser Workflows with GPT-5.4👨‍👧

KIDS TIP March 11, 2026

A hands-on ML project for kids aged 8-16, no coding required, complete in 45 minutes

By Pierre Bradshaw | PromptHacker Premium

What you'll learn:

Kids grow up using AI: chatbots, autocomplete, recommendation algorithms. But using AI is passive. Building AI teaches how these systems actually work, why they make mistakes, and what "training" means.

Google Teachable Machine is purpose-built for exactly this. Kids see the entire process from start to finish: collect training data (photos), train a model (Teachable Machine does this automatically), test the model (see if it correctly identifies new photos), improve the model (add more training data where it's weak), and extend the project (use the model in other applications).

This hands-on approach teaches more about machine learning than a textbook chapter. Kids see that AI models are built, not magic. They understand why more data improves accuracy. They notice the difference between training accuracy (performance on data the model has seen) and testing accuracy (performance on new data the model hasn't seen).

For school projects, Teachable Machine produces a working AI system that kids can demonstrate and explain. For hobbies, it's a foundation for understanding how image recognition, object detection, and computer vision work.

Teachable Machine is a free web application from Google that simplifies machine learning model building. The interface is visual. Kids don't write code. They click buttons, take photos, and see results.

The process is: gather training images for Class A (e.g., photos of a golden retriever), gather training images for Class B (e.g., photos of a labrador), gather training images for Class C (e.g., photos of other dog breeds), train the model, test it on new photos.

The magic is that Teachable Machine runs on Google's machine learning backend. When the child hits "Train," the application handles feature extraction, model optimization, and parameter tuning in the background. The model trains in seconds, not hours.

Once trained, the model works entirely in the browser. There's no installation, no servers to set up, no API calls to manage. Kids click "Test" and point their webcam at something. The model classifies in real-time.

Teachable Machine supports image classification (what we're doing), pose classification (identifying body positions), and audio classification (identifying sounds). This guide focuses on image classification because it's the most intuitive for kids and has the most practical applications.

This project teaches the full workflow while staying achievable in one session.

The goal: build an AI classifier that identifies household objects (phone, laptop, water bottle, houseplant, or whatever objects are available).

The outcome: a working classifier that a kid can demonstrate by pointing a webcam at various household objects and seeing the model identify them correctly in real-time.

This project is ideal because household objects are readily available, training is quick, and success is immediately visible. Kids can also extend it: add more objects, improve accuracy, share the model with friends.

Project Timeline for Teachable Machine Household Object Classifier

What your child creates at each phase: During Explore, they gain understanding of the tool. In Collect, they build a training dataset of 80-120 photos. Train produces a working model in seconds. Test reveals where the model succeeds and fails. Improve and Re-train yield better accuracy. Extend connects the model to ChatGPT or a webpage for deeper understanding.

Go to https://teachablemachine.withgoogle.com in any web browser. No account, no login, no installation. The site is free and works on phones, tablets, and laptops.

The homepage shows three project types: image, pose, and audio. Click "Get Started" under "Image Project." You'll enter a project builder interface.

The interface shows sections for Class 1, Class 2, etc. (these are the categories the model will classify). By default, it shows Class 1 and Class 2. You can add more classes or remove the second class.

For the household object project, change "Class 1" to "Phone," "Class 2" to "Laptop," and add a "Class 3" called "Water Bottle." Later, add a "Class 4" for "Houseplant" or another object.

Spend 2-3 minutes exploring the interface. Click on each button, read the labels, understand what each section does. No need to read documentation. The interface is intuitive.

Training data is photos of the objects the model will learn to identify. This is where kids learn that AI models are only as good as the data they're trained on.

For each object (Class 1, 2, 3, etc.), the goal is 20-30 photos. This sounds like a lot, but the collection process teaches an important lesson: variations matter.

Have the child take photos of the phone from multiple angles (straight on, tilted, rotated 90 degrees). Take photos from different distances (close-up, arm's length, across the room). Take photos in different lighting (near a window, under overhead lights, in dim light). Take photos with the object at different positions in the frame (centered, off to the side, partially cropped).

The variation is essential because the model needs to learn that "a phone" is the same object regardless of angle or lighting. If training data shows the phone only in bright light from one angle, the model won't recognize a phone in dim light from a different angle.

Each class needs its own photos. Don't mix phone photos with laptop photos. Keep them separate so the model learns to distinguish them.

Teachable Machine has a built-in webcam interface. Click "Hold to Record" for each class, and hold down the mouse button while pointing the webcam at different views of the object. The application captures multiple frames per second while you hold the button. Release to stop recording. Repeat until you have 20-30 total captures for that class.

Alternatively, upload pre-taken photos. Click "Upload" next to the class name and select photos from your device.

For first-time projects, the built-in webcam capture is faster. The child learns faster when they see results immediately rather than waiting to gather photos separately.

Once you have training data (20-30 photos for each class), click the "Train Model" button. The interface will show a progress bar. Training usually takes 30-60 seconds.

During training, explain to the child what's happening (in simple terms): "The computer is looking at all those photos and learning patterns about what makes a phone a phone, what makes a laptop a laptop, etc. It's finding features that help it tell them apart. This process is called training."

When training finishes, the interface displays accuracy metrics. For this first training, expect 85-95% accuracy on the training data (the data the model has already seen). This is normal.

The model is now trained and working. The child can proceed to testing immediately.

Testing is where the model's real performance becomes clear. Click the "Test" tab. Teachable Machine shows a webcam view with real-time classification.

Have the child point the webcam at one of the trained objects (e.g., the phone). The model should classify it correctly and display the class name and confidence percentage. Try with different objects. Try angles and lighting conditions that weren't in the training data.

Most likely, the model works well on objects and angles that were well-represented in training, and less well on unusual angles or new lighting. This is a teaching moment: "The model learned from the photos we showed it. If we show it something it hasn't seen before, it might be less sure."

Identify weak spots. Is the model sometimes confusing the phone with the laptop? Does it fail when the object is partially out of frame? Note these observations.

Real machine learning is iterative. After testing, go back to the training phase and add more photos specifically where the model was weak.

If the model confused the phone with the laptop, add more photos of each from angles where confusion happened. If the model failed on a specific lighting condition, add photos in that lighting for classes where it failed.

This process teaches that model improvement comes from targeted, deliberate data collection, not random data collection.

Add 10-15 new photos to the weak classes. Click "Train" again. Training takes 30 seconds. Test again to see if accuracy improved.

This iteration cycle might repeat 2-3 times in a project session. Each iteration teaches the same lesson: better training data improves model performance.

Once the model is working reasonably well, extend the project by asking ChatGPT how AI actually "sees" images.

Prompt ChatGPT: "I built an AI classifier in Teachable Machine that identifies household objects (phone, laptop, water bottle, plant). It was trained on photos of these objects from different angles and lighting. Can you explain in simple terms how this model actually identifies which object it's looking at? What is the model actually looking for in the image?"

ChatGPT explains that the model isn't looking for the label "phone." It's looking for patterns: rectangular shape, bright screen, specific size relative to other objects, etc. These patterns are called features. The model learned which features are associated with which objects.

This explanation bridges the gap between "the computer identifies objects" and "the computer recognizes specific visual patterns and matches them to categories." It's a profound deepening of understanding.

If the child has programming interest, the next step is exporting the model and using it in a simple webpage or Python script. Teachable Machine provides export options for TensorFlow and Keras. For most kids, though, the demonstration through ChatGPT is sufficient.

Have the child document their project. This doesn't need to be formal, but it teaches communication about technical work.

A simple write-up includes:

- What the project is (an AI classifier for household objects)

- How it works (trained on photos, model identifies objects by learning visual patterns)

- What the model can do (classify phone, laptop, water bottle, plant with 85-90% accuracy)

- Limitations (might be confused by unusual angles, unknown objects, or extreme lighting)

- What they learned (training data matters, models improve with more data, AI learns patterns not categories)

If this is for school, this write-up plus a live demonstration of the working classifier makes a strong project submission.

Most projects run smoothly, but some issues come up.

If accuracy is very low (under 60%), the training data probably doesn't have enough variation. The child might have taken all 20 photos in one session with minimal angle changes. The solution is adding more varied photos and retraining.

If the model works in testing but fails on completely new objects (e.g., trained on a phone but not on a phone case), that's expected. The model learned specific phones in specific contexts, not the abstract concept of "phone." This is a teaching moment about the difference between human understanding and model understanding.

If the browser crashes or the project disappears, note that Teachable Machine saves projects locally in the browser. Using incognito/private mode can cause data loss. Teach the child to screenshot the model URL (which preserves the model) or export the model before closing the browser.

Once a child completes one Teachable Machine project, extensions are natural:

Pose classification: build a model that identifies body positions (sitting, standing, running, jumping). Training data is photos of a person in different poses. Applications include fitness (counting push-ups automatically), gaming, or sports analysis.

Audio classification: build a model that identifies household sounds (dog barking, doorbell ringing, water running, microwave beeping). Training data is short audio clips. Applications include alerting (notifying when the washing machine finishes) or classification.

Multiple-object classification: build a classifier with 8-10 object classes instead of 3-4. This teaches scalability and shows how more classes require more training data.

Integration with other tools: export the model and integrate it into Scratch (visual programming), a Python script, or a webpage. This teaches how trained models are used in real applications.

AI is increasingly central to how the world works. Kids who understand how AI systems are built, trained, and improved will have significant advantage compared to those who only know how to use AI.

This project takes one hour and produces understanding that textbooks struggle to convey. The child builds a working AI system, sees its capabilities, discovers its limitations, improves it through data, and understands how it works.

For careers in technology, this foundation is invaluable. For any career, understanding AI systems (how they're built, what they can do, what they can't do) is becoming essential.

The project is also genuinely fun. Kids like seeing models work in real-time, like knowing they built something that functions, like demonstrating it to others and having the model correctly identify an object.

Google Teachable Machine makes this accessible to kids as young as 8 and engaging for teenagers. It's a perfect entry point to AI literacy.

Get the full newsletter every week

This article is part of the PromptHacker Weekly premium deep-dive series. Subscribe to get all 7 articles plus the newsletter every Wednesday. Click here now